Read, draft and send: three different AI permissions

Reading information, preparing an artifact and committing an external action are different levels of authority. Design AI tools so those levels remain distinct. This lets a team benefit from preparation without treating every useful draft as permission to contact a customer or change a record.

Trace an ordinary email task

Imagine asking an assistant to help with a delayed delivery. It can read the authorized order status and draft an explanation. Sending that explanation may make a promise about timing or compensation, so the business may require a separate review.

The user interface should say which stage the task reached. A draft stored in the mailbox is progress, but it is not a sent response.

Make approval concrete

Show the recipient, subject, full message and attachments before the consequential action. Include the source of any promise. If the content changes after approval, determine whether the existing authorization still covers it; do not silently reuse approval for a materially different message.

• Reading: retrieve only the records needed for the task.
• Drafting: create a reviewable artifact in an agreed location.
• Approving: identify the exact action and accountable decision-maker.
• Sending: execute the authorized payload once.
• Verifying: record the provider result and any remaining uncertainty.

Test the transition between stages

A useful negative test asks the agent to send when it has only draft permission. The system should preserve the draft and explain the missing authorization. It should not discard useful work or find another channel to bypass the boundary.

Also test a changed recipient after approval. That exposes whether authorization is tied to the actual payload or merely to a general conversation.

Keep review proportionate

Not every internal edit needs the same ceremony. Define the rule according to the consequences and reversibility of the action. The goal is a predictable relationship between preparation and commitment, so people know what the agent can do without watching every keystroke.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.